3j>SrSSKrSSKJrJrJrJr SSKrSSKJ r SSK J r J r J r SSKJr SSKJr SS KJr SS KJr SS KJr SS KJr S SKJr "SS\\5rg)z1 Ernie-Image Pipeline for HuggingFace Diffusers. N)CallableListOptionalUnion)Image) AutoModelAutoModelForCausalLM AutoTokenizer)ErnieImageLoraLoaderMixin)AutoencoderKLFlux2)ErnieImageTransformer2DModel)DiffusionPipeline)FlowMatchEulerDiscreteScheduler) randn_tensor)ErnieImagePipelineOutputc"^\rSrSrSrSrSS/rS/rS5S\S \ S \ S \ S \ S\ \S\ \ 4U4S jjjr\S5r\S5r\R("5S6S\S\R,S\S\S\ \S\S\S\4Sjj5r\R("5S7S\\\\4S\R,S\S\\R84Sjj5r\S\R8S\R84Sj5r\S\R8S\R84Sj5r \S\\R8S\R,S \RBS!\4S"j5r"\R("5SS#SSS$S%SSSSSS&S'SS/S'4S\ \\\\4S(\ \\\\4S\S\S)\S*\S\S+\ \RFS\ \R8S,\$\RJS-S-\$\RJS-S.\S/\&S0\ \'\\\(/S4S1\\S2\&4 S3jj5r)S4r*U=r+$)8ErnieImagePipeline#a Pipeline for text-to-image generation using ErnieImageTransformer2DModel. This pipeline uses: - A custom DiT transformer model - A Flux2-style VAE for encoding/decoding latents - A text encoder (e.g., Qwen) for text conditioning - Flow Matching Euler Discrete Scheduler z"pe->text_encoder->transformer->vaepe pe_tokenizerlatentsN transformervae text_encoder tokenizer schedulerc >[TU]5 URUUUUUUUS9 [USS5(a2S[ UR R R5-UlgSUlg)N)rrrrrrrr) super__init__register_modulesgetattrlenrconfigblock_out_channelsvae_scale_factor) selfrrrrrrr __class__s n/home/wildlama/miniconda3/lib/python3.13/site-packages/diffusers/pipelines/ernie_image/pipeline_ernie_image.pyr#ErnieImagePipeline.__init__3su  #%%  SZZ^`egkRlRlc$((//*L*L&M NrtcUR$)N_guidance_scaler*s r,guidance_scale!ErnieImagePipeline.guidance_scaleIs###r.c URS:$)N?r0r2s r,do_classifier_free_guidance.ErnieImagePipeline.do_classifier_free_guidanceMs##c))r.promptdevicewidthheight system_prompt temperaturetop_preturnc `[R"XUS.SS9n/n UbU RSUS.5 U RSUS.5 URR U SSS9n URU S S 9R U5n UR R"S0U DURRUS :g=(d US :gUUURRURRS .D6n U S U SRSSn URRU SS9R5$)zAUse PE model to rewrite/enhance a short prompt via chat_template.)r:r<r=F) ensure_asciiNsystem)rolecontentuser)tokenizeadd_generation_promptpt)return_tensorsr6)max_new_tokens do_sampler?r@ pad_token_id eos_token_idr input_idsrT)skip_special_tokens)jsondumpsappendrapply_chat_templatetorgeneratemodel_max_lengthrNrOshapedecodestrip)r*r:r;r<r=r>r?r@ user_contentmessages input_textinputs output_ids generated_idss r,_enhance_prompt_with_pe*ErnieImagePipeline._enhance_prompt_with_peQsAzz @   $ OOX-H ILAB&&:: "'; "":d"CFFvNWW%%  ,,==!S(8ESL#**77**77 #1 f[&9&?&?&B&DE   '' 4'PVVXXr.rnum_images_per_promptc [U[5(aU/n/nUHnURUSSSS9Sn[U5S:Xa2URRbURR/nOS/n[ R "U/US9n[ R"5 URUSS9nURS Sn SSS5 [U5Hn URW 5 M M U$!,(df  N7=f) z"Encode text prompts to embeddings.TF)add_special_tokens truncationpaddingrPrN)r;)rPoutput_hidden_states) isinstancestrrr& bos_token_idtorchtensorno_gradr hidden_statesrangerU) r*r:r;re text_hiddenspidsrPoutputshidden_s r, encode_prompt ErnieImagePipeline.encode_prompt{s fc " "XF A..#' !  C3x1}>>..:>>667C#C cU6:I++')-, !..r215 !01##F+216!s !#C55 D cURupp4URXUS-SUS-S5nURSSSSSS5nURXS-US-US-5$)z12x2 patchify: [B, 32, H, W] -> [B, 128, H/2, W/2]r rrr )rZviewpermutereshaperbchws r,_patchify_latents$ErnieImagePipeline._patchify_latentssf]] a,,qQ!VQQ://!Q1a3qa%aa88r.cURupp4URXS-SSX45nURSSSSSS5nURXS-US-US-5$)z5Reverse patchify: [B, 128, H/2, W/2] -> [B, 32, H, W]r~r rrr}r )rZrrrs r,_unpatchify_latents&ErnieImagePipeline._unpatchify_latentss`]] a//!!VQ18//!Q1a3qq&!a%Q77r.rtdtype text_in_dimc[U5nUS:Xa<[R"SSU4XS9[R"SU[RS94$UVs/sHeoUR 5S:Xa/UR S5R U5R U5OUR U5R U5PMg nn[R"UVs/sHowRSPM snU[RS9n[UR5R55n [R"XIU4XS9n [U5HupXzU SURS2SS24'M X4$s snfs snf)Nrr;r)rr r) r&rozeroslongdimsqueezerWrprZintmaxitem enumerate) rtr;rrBth normalizedtlensTmaxtext_bthis r, _pad_textErnieImagePipeline._pad_textsI   6;;1k26OQVQ\Q\V5::R ht gsac&&(a-BJJqM  V $ ' ' .RUU6]EUEUV[E\ \gs  ||<AWWQZU b(U VVs/sHn[U5HnUPM M nnnOURUUU5nX0R -nX@R -nURR R"nU c[%UUUU4UUUS 9n [&R("S S US -5nUR*R-USSUS9 UR(a[W5[U5-nOUnUR/UUUURR R0S9un n!UR3US9n"[5UR*R65GH7un#n$UR(a@[&R8"X/SS9n%[&R:"US-4U$R=5UUS9n&O(U n%[&R:"U4U$R=5UUS9n&URU%U&U U!SS9Sn'UR(aU'R?SSS9un(n)U(UU)U(- --n'UR*RAU'U$U 5RBn Ub80n*UHn+[E5U+U*U+'M U"UU#U$U*5n,U,RGSU 5n U"RI5 GM: SSS5 U S:XaU $URJRLRNRQS SS S 5RSU5n-[&RT"URJRLRVRQS SS S 5S-5RSU5n.U U.-U--n URYU 5n URJR[U SS9Sn/U/R]SS 5S -S- n/U/R_5RaSSSS 5Rc5Re5n/U S:Xa8U/V0s/sH+n0[fRh"U0S-RkS55PM- n/n0URm5 U (dU/4$[oU/US9$s snfs snnfs snnf!,(df  GN=fs sn0f) al Generate images from text prompts. Args: prompt: Text prompt(s) negative_prompt: Negative prompt(s) for CFG. Default is "". height: Image height in pixels (must be divisible by 16). Default: 1024. width: Image width in pixels (must be divisible by 16). Default: 1024. num_inference_steps: Number of denoising steps guidance_scale: CFG scale (1.0 = no guidance). Default: 4.0. num_images_per_prompt: Number of images per prompt generator: Random generator for reproducibility latents: Pre-generated latents (optional) prompt_embeds: Pre-computed text embeddings for positive prompts (optional). If provided, `encode_prompt` is skipped for positive prompts. negative_prompt_embeds: Pre-computed text embeddings for negative prompts (optional). If provided, `encode_prompt` is skipped for negative prompts. output_type: "pil" or "latent" return_dict: Whether to return a dataclass callback_on_step_end: Optional callback invoked at the end of each denoising step. Called as `callback_on_step_end(pipeline, step, timestep, callback_kwargs)` where `callback_kwargs` contains the tensors listed in `callback_on_step_end_tensor_inputs`. The callback may return a dict to override those tensors for subsequent steps. callback_on_step_end_tensor_inputs: List of tensor names passed into the callback kwargs. Must be a subset of `_callback_tensor_inputs` (default: `["latents"]`). use_pe: Whether to use the PE model to enhance prompts before generation. Returns: :class:`ErnieImagePipelineOutput` with `images` and `revised_prompts`. Nz0Must provide either `prompt` or `prompt_embeds`.zBCannot provide both `prompt` and `prompt_embeds` at the same time.rz&Height and width must be divisible by )r<r=rz1negative_prompt must have same length as prompt ())rr;rr6gr)sigmasr;)rtr;rr)total)rr rF)rrtimestepr text_lensrrlatentgh㈵>)rr ruint8)imagesrevised_prompts)8_execution_devicerrr1 ValueErrorr)rlrmrrrclistr&rsrzr7r' in_channelsrrolinspacer set_timestepsrr progress_barr timestepscatfullrchunkstep prev_samplelocalspopupdaterbn running_meanrrWsqrt running_varrr[clampcpurfloatnumpyr fromarrayastypemaybe_free_model_hooksr)1r*r:rr=r<rr3rerrrrrrrrrr;rrru batch_sizetotal_batch_sizerryrtuncond_text_hiddenslatent_hlatent_wlatent_channelsrcfg_text_hiddensrrrrrlatent_model_inputt_batchpred pred_uncond pred_condcallback_kwargskcallback_outputsbn_meanbn_stdrimgs1 r,__call__ErnieImagePipeline.__call__sd''  &&- >m3OP P  -";ab b )) )Q .%:O:O2OST2TEdF[F[E\]^ ^  &#&& 04  &TWW-@TEVEVEbcijci^_d221fE2YciFj"6lO  VJ]+J%(==  " O os + +./*ZA>ZA}L--ff>STL  + +%12H&o2HQRWXmRnQqRnq2H#&o#&*&8&8&Rg&h#222111**11== ?"!?HhG# GS*=*AB $$F3BK$G  + +#$784 ;MM + "nn)&SWScScSjSjSvSv- )  %8  9\!$..":":;133).G3E1)M&#jj*:Q*>)@!&&(SYafgG)0&#jj*:): 9pVs% WW (WEW-2W( W%)r1r))NN)r9r9Ng333333?gffffff?)r),__name__ __module__ __qualname____firstlineno____doc__model_cpu_offload_seq_optional_components_callback_tensor_inputsrr rr rrr r#propertyr3r7rorqrmr;rrrcrrTensorrz staticmethodrrrr Generatorr FloatTensorboolrdictr__static_attributes__ __classcell__)r+s@r,rr#sA .1(k.204u1u u u ! u 3 u ) *u}-uu,$$** ]]_ '+ 'Y'Y 'Y 'Y  'Y  } 'Y'Y'Y 'Y'YR ]]_ &' 'c49n%' ' # ' ell  ''R95<<9ELL998U\\8ell88U\\ 2ELLQVQ\Q\kn  ]]_37;=#% #%&/3*.8AXAXAX'xc40@$0F'GHAX -1I!AX"#AXAXr.r)rrStypingrrrrroPILr transformersrr r loadersr modelsr models.transformersrpipelines.pipeline_utilsr schedulersrutils.torch_utilsrpipeline_outputrrrRr.r,rsJ 22 GG0(?99-5eX*,EeXr.